Yep. I use an Amazon EC2 Linux AMI with nVidia GRID GPU's using CUDA and OpenCV and it works like a charm. Of course it is headless so I tunnel X11 through ssh and use xv to look at the images I am rendering. Saves me from buying hardware. Costs me $0.68 an hour -- well worth it.

They don't mention any manufacturers and do not support CUDA, just OpenGL.

It's clear what they are doing: you call Amazon's OpenGL library, which applies some batching and compression when talking to a remote GPU somewhere else in the cluster. You are not allowed to or even need to know what kind of hardware is on the other side. They could even pick different manufacturers. And because of this proxying, you can only use an open and vendor-neutral API like GL, hence no CUDA support.